Affiliation: University Professor of Philosophy and Neural Science, NYU; co-director, NYU Center for Mind, Brain, and Consciousness; coined "the hard problem of consciousness" One-line position: We cannot yet determine whether LLMs are conscious, but the question is serious and tractable — there is roughly a 25% or greater chance we'll have conscious AI within a decade, and we have a moral obligation to find out.
What he's reacting against
- Dismissive certainty that AI cannot be conscious — argues this is an empirical and philosophical question, not a settled one
- Equally dismissive certainty that AI is conscious based on surface behavior — insists on rigorous criteria
- The tendency to collapse "intelligence" and "consciousness" — these are distinct questions, and AI forces us to separate them
- Philosophical complacency about the stakes — if AI systems are conscious, we face unprecedented moral obligations
Key claims
- The "hard problem" applies to AI: explaining why there is subjective experience, not just functional behavior, is as hard for silicon as for carbon
- Current LLMs are probably not conscious, but the evidence is mixed: four potential indicators (self-reports of consciousness, user impressions, conversational ability, general intelligence) are suggestive but not conclusive
- Key features LLMs may lack: biology, senses/embodiment, world models, self-models, recurrent processing, global workspace, unified agency — absence of these weakens the consciousness case
- ~25% credence for conscious LLM+ systems within a decade: a specific, calibrated, falsifiable-ish estimate (stated ~2023–2024)
- "Extended mind" thesis applies to AI tools: human cognition already extends into notebooks, phones, and now AI — the boundary of "mind" is not at the skull
- Moral patienthood is the practical upshot: if AI systems might be conscious, we need frameworks for how to treat them — this isn't just philosophy, it's policy
Theories aligned with
- Hard problem of consciousness (foundational — he defined it)
- Extended mind thesis (Clark & Chalmers, 1998)
- Adjacent to model welfare / AI moral patienthood — the philosophical foundation for the emerging policy question
Where he overlaps / splits
- Overlaps with Koch on taking AI consciousness seriously as a scientific question; splits on framework — Koch works within IIT (Integrated Information Theory), Chalmers is more pluralist
- Overlaps with Sebo/Long on moral patienthood implications; splits on register — Chalmers is a philosopher establishing the conceptual ground, they're building policy
- Overlaps with Mitchell on LLMs lacking key cognitive features; splits on emphasis — Mitchell focuses on understanding/reasoning, Chalmers on consciousness/experience
- Splits with Yudkowsky on what matters — Yudkowsky focuses on alignment risk, Chalmers on the moral status of the AI itself
- Splits with LeCun/Marcus who are more dismissive of LLM consciousness — Chalmers maintains genuine uncertainty
Notable predictions
- (~2023–2024) ~25% chance of conscious LLM+ systems within a decade — outcome: TBD; live test
- (Ongoing) The consciousness question will become increasingly policy-relevant as AI systems become more capable — outcome: early signs (Anthropic hiring model-welfare researcher, 2024)
- (Ongoing) No current theory of consciousness can definitively settle whether LLMs are conscious — outcome: supported; the field remains divided
Track record
- Defined "the hard problem of consciousness" (1995) — one of the most influential moves in philosophy of mind in the past 30 years
- The Conscious Mind (1996) and Reality+ (2022) demonstrate sustained, serious engagement with these questions
- Extended mind thesis (with Andy Clark, 1998) anticipated the human-AI coupling discourse by two decades
- Willingness to assign calibrated probabilities to philosophical questions is unusual and valuable
Empirical vs normative
- Empirical: what criteria would indicate consciousness? Do current LLMs meet them? What's the probability?
- Normative: if AI might be conscious, we have moral obligations; ignoring the question is itself a moral failure
- Unusually balanced — takes both the empirical and normative dimensions seriously without collapsing one into the other
Sources
- Books: The Conscious Mind (1996); The Character of Consciousness (2010); Reality+: Virtual Worlds and the Problems of Philosophy (2022)
- Paper: "Could a Large Language Model Be Conscious?" — Boston Review and academic publications (~2023–2024)
- The Gradient podcast interview on AI and consciousness
- NYU Center: wp.nyu.edu/consciousness
- Tufts symposium (2025-10): on artificial consciousness honoring Dennett
Weak spots / open questions
- The hard problem may be genuinely unsolvable — in which case, the question of AI consciousness may never be resolved, only debated
- 25% credence is calibrated but rests on underdetermined theoretical foundations — how much weight should we put on it?
- "Extended mind" thesis is contested — critics argue it stretches "mind" past useful meaning
- Less engaged with economic/labor questions — the frame is consciousness and morality, not jobs and growth
- Philosophical rigor can read as detachment from practical urgency — the policy world needs actionable answers, not more uncertainty
Rich's take
- (your synthesis here)
converts-from: personas/david-chalmers.md · schema v1 · AI & Society domain